Scalable Multi-User Precoding and Pilot Optimization with Graph Neural Networks

We consider the problem of sum rate maximization in frequency division duplex (FDD) systems when only imperfect channel state information (CSI) is available. Inspired by the low complexity and generalization ability offered by graph neural networks (GNNs), we propose an end-to-end (E2E) framework for both, precoding and downlink (DL) pilot sequences optimization based on the novel Edge-graph attention network (GAT). The simulation results confirm the potential of the proposed E2E approach to optimize sum rates and its scalability in scenarios with varying numbers of users. Additionally, the superiority of the learned pilot matrix compared to the conventionally employed sub-discrete Fourier transform (DFT) matrix is highlighted.